Folding Forum

Rattledagger wrote:Under BOINC you don't need a separate SMP-client, this is a limitation of FAH's clients that is finally fixed with FAH's v7-client.

FAH needed to beta an upgraded client that included the add-on MPI libraries for Windows. It wasn't really a "separate" client, but rather an extension of the standard client. At about the same time, they did a beta of a client containing the support files for the GPU cores. Thus began a saga into a standard client plus two beta branches. At any time, the SMP branch could have been called "released" and could have replaced the standard client, but the MPI libraries were never robust enough for that to happen.

V7 is a comprehensive replacement that manages run-time libraries for the older ati and nvidia cores as well as newer ones for for nvidia, for opencl. I don't think there are any discrete libraries for mpi any more and the gpu libraries may have just been incorporated into the cores or the driver packages. One if the disadvantages of being an early-adopter is that you have to live through the growing pains of incompatible versions and the questions of who is responsible for the installation process before the details are smoothed out. All I know is that each one works fine now.

gwildperson wrote:

Rattledagger wrote:Under BOINC you don't need a separate SMP-client, this is a limitation of FAH's clients that is finally fixed with FAH's v7-client.

FAH needed to beta an upgraded client that included the add-on MPI libraries for Windows. It wasn't really a "separate" client, but rather an extension of the standard client.

Well, having to run different clients in separate directories atleast to me fits my description...

Was also thinking about the Wikipedia-description:

The project differs from other distributed computing projects, such as those under BOINC, by offering a variety of clients such as those for multi-core processors, graphics processing units, and PlayStation 3s, complementing the standard client designed for uniprocessor systems

Except for the PS3-part, under BOINC a single client-install handles all these features, you don't need multiple clients in separate directories to handle this and you don't need to fool-around with separate machineid's.

Well that passage in the article is completely my words, because I was considering "clients" as meaning both applications and cores for the purposes of project-project comparison. If any DC project, BOINC or not, launched some software that uses SMP for computing before Nov 2006, then we are not the first. Perhaps the more general term "software" is more appropriate.

EDIT: The comparison in the passage is no longer valid. At the time, I didn't know that BOINC had more than just uniprocessors. I'll shortly be removing that comparison entirely.

EDIT 2: Comparison removed.

As was pointed out previously, SMP isn't a useful technique when a project has plenty of independent assignments. The only reason FAH needs it is because it has assignments (trajectories) that will take dozens of years to complete using a single processor. SMP adds some overhead, but it also dramatically reduces the total elapsed time by using multiple CPUs to work on the same assignment.

I'm still confused about how any of the *@home projects first used SMP and why. (BOINC is not a single *@home project, but rather a tool that allows processing to be shared across multiple *@home type projects.) If aqua@home used SMP, why did they need to accelerate the processing of single WUs?

codysluder wrote:I'm still confused about how any of the *@home projects first used SMP and why. (BOINC is not a single *@home project, but rather a tool that allows processing to be shared across multiple *@home type projects.) If aqua@home used SMP, why did they need to accelerate the processing of single WUs?

Another quote by Kamran Karimi tells this:

This seems to be more complicated than I imagined, so I'll leave it as it is now. The reason I asked the question is that AQUA's work units now take weeks to complete on a CPU, but "only" days on a GPU, so I was looking for a way to have more GPU owners join the project.

-Kamran

Source, 29. May 2009, http://www.archivum.info/boinc_projects ... sible.html

Of course, Climateprediction.net at the same time ran wu's taking weeks to months to finish, so that Aqua@home "needed" SMP can be argued...

By it's very nature, Gromacs needed both SMP and GPU support, using them to solve different types of problems Both are, as he says, complex. Fortunately both of those developments were open-sourced as part of gromacs.org and openmm.org projects for the benefit of scientists everywhere, as well as being used within FAH.